TEBD algorithm fails to converge

[manthan-badbaria]

Hi,
I need to evolve an MPS state(pure state) in real time. For the same I am using TEBD function provided by quimb.
My parameters are as follows:
tebd.split_opts['cutoff']=1e-12 # I require large accuracy
tol=1e-7
I evolve different pure states in MPS form using TEBD but for some states I get the following issue:
"Internal algorithm failed to converge".
Just to mention I have played around with various values of tolerance and cutoff but I end up getting this issue. Also the same calculation worked fine months back.

[nikitn2]

Yes, I also get a lot of "Internal algorithm failed to converge" errors when using the .compress() function. Did something change between update?

EDIT: I'm using Mac's accelerate framework for the M1 Max CPU. Might this be an issue?

[jcmgray]

Yes, along with #189, I wonder if something has changed about the numba svd implementation. I don't recall having changed anything on the quimb side of things. Repeating my request there, it would be helpful to see if:

1. it appears for different version of numba going back
2. it appears for a different backend like torch.
3. you could save an actual matrix/tensor it occurs on.

[nikitn2]

Hi Johnnie,

Thank you for replying.

My numba version is 0.56.4, from back in Nov 2022, so I don't know if it's the issue. Only real change I've done to my environment is to use the Apple “accelerate” framework for BLAS, which I've heard can be more unstable than openblas. However, since several others are apparently sufering from similar problems, Apple might not be to blame for this one unless the others also happen to use accelerate-BLAS.

What I do notice however, is that the eig & isvd methods seem more stable than svd and svds. I wonder if renorm = True might be helpful? I did some tests with this, but can't say anything too conclusive.

However, I tested the workaround in #189 where the scipy-svd is used as a fall-back, and it works quite well for me.

[manthan-badbaria]

Hi @nikitn2 ,
I tried the work around in #189 by monkey patching the function with the new modified function. But as the function is not defined in some class, it does not work for me. Can you share some basic code of how you did it?

[nikitn2]

Hi @manthan-badbaria,

Yes, a package is missing which you need to import; I think it was something from Scipy, perhaps scipy.linalg. Just google the function that's missing and import it.

[manthan-badbaria]

Hi Johnnie, Thanks for your reply. I am sharing the error msg I get if that helps, also my numba version is 0.56.4 Traceback (most recent call last): File "/Users/manthanbadbaria/anaconda3/lib/python3.10/site-packages/joblib/externals/loky/process_executor.py", line 436, in _process_worker r = call_item() File "/Users/manthanbadbaria/anaconda3/lib/python3.10/site-packages/joblib/externals/loky/process_executor.py", line 288, in __call__ return self.fn(*self.args, **self.kwargs) File "/Users/manthanbadbaria/anaconda3/lib/python3.10/site-packages/joblib/_parallel_backends.py", line 595, in __call__ return self.func(*args, **kwargs) File "/Users/manthanbadbaria/anaconda3/lib/python3.10/site-packages/joblib/parallel.py", line 263, in __call__ return [func(*args, **kwargs) File "/Users/manthanbadbaria/anaconda3/lib/python3.10/site-packages/joblib/parallel.py", line 263, in <listcomp> return [func(*args, **kwargs) File "/var/folders/dw/9qb0srzj4zs78hy3wnh069mw0000gn/T/ipykernel_34242/288811871.py", line 17, in Evolution File "/Users/manthanbadbaria/anaconda3/lib/python3.10/site-packages/quimb/tensor/tensor_1d_tebd.py", line 438, in at_times self.update_to(t, dt=dt, tol=False, order=order, progbar=False) File "/Users/manthanbadbaria/anaconda3/lib/python3.10/site-packages/quimb/tensor/tensor_1d_tebd.py", line 391, in update_to self.step(order=order, progbar=progbar, dt=None, queue=True) File "/Users/manthanbadbaria/anaconda3/lib/python3.10/site-packages/quimb/tensor/tensor_1d_tebd.py", line 333, in step {2: self._step_order2, File "/Users/manthanbadbaria/anaconda3/lib/python3.10/site-packages/quimb/tensor/tensor_1d_tebd.py", line 328, in _step_order4 self._step_order2(tau1, **sweep_opts) File "/Users/manthanbadbaria/anaconda3/lib/python3.10/site-packages/quimb/tensor/tensor_1d_tebd.py", line 316, in _step_order2 self.sweep('left', tau, **sweep_opts) File "/Users/manthanbadbaria/anaconda3/lib/python3.10/site-packages/quimb/tensor/tensor_1d_tebd.py", line 266, in sweep self._pt.gate_split_( File "/Users/manthanbadbaria/anaconda3/lib/python3.10/site-packages/quimb/tensor/tensor_1d.py", line 1608, in gate_split return self.gate_inds( File "/Users/manthanbadbaria/anaconda3/lib/python3.10/site-packages/quimb/tensor/tensor_core.py", line 3488, in tensor_network_gate_inds _tensor_network_gate_inds_basic( File "/Users/manthanbadbaria/anaconda3/lib/python3.10/site-packages/quimb/tensor/tensor_core.py", line 3198, in _tensor_network_gate_inds_basic tln, *maybe_svals, trn = tlGr.split( File "/Users/manthanbadbaria/anaconda3/lib/python3.10/site-packages/quimb/tensor/tensor_core.py", line 2231, in split return tensor_split(self, *args, **kwargs) File "/Users/manthanbadbaria/anaconda3/lib/python3.10/site-packages/quimb/tensor/tensor_core.py", line 514, in tensor_split left, s, right = _SPLIT_FNS[method](array, **opts) File "/Users/manthanbadbaria/anaconda3/lib/python3.10/site-packages/autoray/autoray.py", line 834, in __call__ return fn(*args, **kwargs) File "/Users/manthanbadbaria/anaconda3/lib/python3.10/site-packages/numba/np/linalg.py", line 879, in _handle_err_maybe_convergence_problem raise ValueError("Internal algorithm failed to converge.") ValueError: Internal algorithm failed to converge. It's very tricky to reproduce it even using the same parameters

…

On Fri, Jul 14, 2023 at 7:40 PM nikitn2 ***@***.***> wrote: Hi Johnnie, Thank you for the reply. My numba version is 0.56.4, from back in Nov 2022, so I don't know if it's the issue. Only real change I've done to my environment is to use the Apple “accelerate” framework for BLAS, which I've heard can be more unstable than openblas. However, since several others are apparently sufering from similar problems, Apple might not be to blame for this one unless the others also happen to use accelerate-BLAS. What I do notice however, is that the eig & isvd methods seem more stable than svd and svds. I wonder if renorm = True might be helpful? I did some tests with this, but can't say anything too conclusive. However, I tested the workaround in #189 <#189> where the scipy-svd is used as a fall-back, and it works quite well for me. — Reply to this email directly, view it on GitHub <#192 (reply in thread)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AWKXFJS7NMSVNXN4NJCCKV3XQFHMLANCNFSM6AAAAAA2HKJEHA> . You are receiving this because you authored the thread.Message ID: ***@***.***>